Yunpeng Jiang

I'm a PhD student at UM-SJTU Joint Institute in Shanghai, working on sample-efficient deep reinforcement learning. I am currently advised by Yutong Ban and Paul Weng. Before my PhD studies, I completed my bachelor's degree in electrical and computer engineering at the same institute.

Email  /  Scholar  /  Github

Research

I'm interested in deep reinforcement learning and robot manipulation. My current research focuses on improving sample-efficiency of robot learning algorithms by exploiting data augmentation and leveraging equivariance. Some papers are highlighted.

We propose Time Reversal symmetry enhanced Deep Reinforcement Learning (TR-DRL), a framework that combines trajectory reversal augmentation and time reversal guided reward shaping to efficiently solve temporally symmetric tasks.

We reformulate classification using a novel variant of fair optimization and propose a multiplicative weight optimization method to reduce the class-dependent effects of data augmentation.

We make recommendations on how to exploit data augmentation in image-based DRL in a more principled way. And we include a novel regularization term called tangent prop in RL training.

We exploit a recent neuro-symbolic model called differentiable logic machine to learn an interpretable controller in the form of a first-order logic program. We demonstrate the feasibility of our approach on two classical decision-making scenarios in autonomous driving: lane changing and intersection management.

This website is build on this source code. Thanks to Jon Barron.